This research demonstrates how Low-Rank Adaptation (LoRA) significantly reduces unintended memorization of sensitive patient data when training medical LLMs in federated learning environments.

• Privacy Vulnerability: Standard federated learning still allows LLMs to memorize and potentially leak sensitive medical information
• Novel Solution: LoRA fine-tuning reduces memorization by up to 33% while maintaining model performance
• Practical Implementation: Provides a framework for more private collaborative medical AI training without sacrificing utility
• Technical Innovation: Shows how parameter-efficient tuning methods can have privacy benefits beyond computational efficiency

For healthcare organizations, this approach enables safer collaboration on AI models across institutions while better protecting patient data confidentiality and regulatory compliance.

Mitigating Unintended Memorization with LoRA in Federated Learning for LLMs